Frozen Feature Augmentation for Few-Shot Image Classification

Training a linear classifier or lightweight model on top of pretrained visionmodel outputs, so-called 'frozen features', leads to impressive performance ona number of downstream few-shot tasks. Currently, frozen features are notmodified during training. On the other hand, when networks are trained directlyon images, data augmentation is a standard recipe that improves performancewith no substantial overhead. In this paper, we conduct an extensive pilotstudy on few-shot image classification that explores applying dataaugmentations in the frozen feature space, dubbed 'frozen feature augmentation(FroFA)', covering twenty augmentations in total. Our study demonstrates thatadopting a deceptively simple pointwise FroFA, such as brightness, can improvefew-shot performance consistently across three network architectures, threelarge pretraining datasets, and eight transfer datasets.